Selecting Near-Optimal Learners via Incremental Data Allocation